HELM Architecture

Architecture diagrams for the HELM boundary.

Start with the execution boundary, then browse how company work, draft specs, PEP/CPI checks, receipts, ProofGraph records, and EvidencePacks connect.

Define the architecture terms

CPI: The check that validates a plan before HELM lets it continue.
PEP: The boundary that enforces the decision before a tool call runs.
ProofGraph: A record chain that helps replay and check what happened.
EvidencePack: A small bundle of records used to verify one event or review path.

Boundary

Execution Boundary

Shows where model proposals become governed HELM decisions before side effects.

Inputs: Actor, proposed action, policy, context, approval state.
Outputs: ALLOW, DENY, or ESCALATE plus receipt and proof pointer.
Risk: Without this boundary, valid-looking agent output can still dispatch unsafe work.
Proof artifact: Receipt, policy verdict, signed intent, replay path.

Read the docs page

Operations

Autonomous Loop

Connects company context, proposal, review, execution, and proof into one operating loop.

Inputs: Company artifacts, policies, tickets, approvals, agent proposals.
Outputs: Reviewed work requests and governed side effects.
Risk: A loop without proof becomes automation theater rather than accountable operation.
Proof artifact: Receipts written back to proof and company context surfaces.

Read the docs page

Proof

ProofGraph Trail

Explains how decisions remain replayable after the action has run or been blocked.

Inputs: Verdicts, receipts, approvals, policy snapshots, effect records.
Outputs: Checkable proof trail for operators, security teams, and auditors.
Risk: If proof is missing, later review depends on recollection and logs instead of a bounded evidence path.
Proof artifact: ProofGraph records and EvidencePack references.

Read the docs page

Proof

Proof Condensation

Shows how evidence can be reduced into reviewable packs without losing decision context.

Inputs: Receipt streams, policy snapshots, approval records, replay metadata.
Outputs: Smaller EvidencePack slices for audits and reviews.
Risk: Over-condensing proof can hide the reason a decision was allowed, denied, or escalated.
Proof artifact: EvidencePack slice with references back to source records.

Read the docs page

Compiler

OrgGenome Compiler

Frames compiler output as draft material until VGL review, simulation, P0 ceilings, approval, attestation, and receipts promote it.

Inputs: OrgDNA, policies, interviews, manifests, examples, existing process records.
Outputs: Draft rules, process graphs, activation candidates, mirror hashes, impact reports.
Risk: Treating generated rules as authority before activation collapses the HELM boundary.
Proof artifact: Semantic mirror, signed OrgGenome hash, P0 ceiling binding, approval attestation, activation receipt.

Read the docs page

Company Truth

Code Intelligence Graph

Shows how source-backed code reality informs engineering specs without becoming execution authority.

Inputs: Repo commit, file hashes, symbols, routes, call edges, tests, CodeIndexReceipt.
Outputs: CodeImpact, read_set, write_set, affected tests, route impact, closure evidence.
Risk: Engineering agents can exceed approved scope if code evidence is stale, missing, or treated as authority.
Proof artifact: CodeIndexReceipt, CodeImpact report, test receipts, PR refs, EvidencePack.

Read the docs page

Company Truth

Queryable -> Executable

Separates reading company state from executing company action.

Inputs: Source-backed company context and proposed work.
Outputs: A reviewed execution path only after HELM checks the action.
Risk: Queryable context can be useful while still being unsafe as execution authority.
Proof artifact: Source references, CPI verdict, receipt.

Read the docs page

Company Truth

Should-vs-Is

Explains how HELM AI Enterprise can surface gaps between expected and observed company state.

Inputs: Roadmaps, tickets, commitments, PRs, incidents, receipts.
Outputs: Truth conflicts, drift signals, and candidate work requests.
Risk: A detected gap is not permission to act; it is a reason to review and gate the next step.
Proof artifact: Source-linked conflict and any later execution receipt.

Read the docs page

Operations

Product Surface Map

Maps public, docs, HELM AI Kernel, HELM AI Enterprise, and trust surfaces without merging their authority.

Inputs: Product routes, docs routes, proof surfaces, contact paths.
Outputs: Clear navigation from education to evaluation to architecture review.
Risk: Public surfaces can overreach if they blur HELM AI Kernel, HELM AI Enterprise, private pressure tests, and research.
Proof artifact: Trust Center, docs links, and source details.

Read the docs page

Company Truth

Digital Twin

Contrasts company state visibility with the separate authority required for action.

Inputs: Company graph, source artifacts, state projections.
Outputs: Readable context for planning and review.
Risk: A mirror of the company is not the right place to grant side-effect authority.
Proof artifact: Source-backed graph references plus gated execution records.

Read the docs page

Connectors

Digital-Gateway Bridge

Shows why analog and kinetic effects need tighter limits, safety contracts, telemetry, and often escalation.

Inputs: Vendor, shipment, field-service, AMR mission, factory workflow, or device-gateway requests.
Outputs: Bounded action, denied path, or escalated command-gateway proposal.
Risk: Real-world-adjacent effects are harder to roll back than software state.
Proof artifact: Approval, contract, safety profile, telemetry receipt, jurisdiction boundary, EvidencePack.

Read the docs page

Boundary

Authority Layer

Positions HELM as the authority layer between model output and real systems.

Inputs: Planner output, user identity, tenant, policy, connector scope.
Outputs: Deterministic decision and enforceable permission boundary.
Risk: Self-policing agents can confuse fluent output with authorization.
Proof artifact: Policy verdict, sandbox grant, receipt.

Read the docs page

Operations

Research Library

Organizes research concepts without making long-horizon ideas current product copy.

Inputs: Public research pages and product status.
Outputs: Reader paths into current, gated, and non-normative material.
Risk: Research language can dilute the execution-boundary wedge if product status is not explicit.
Proof artifact: Source details and status labels.

Read the docs page